1. 15.053Tuesday, April 9 Branch and Bound Handouts: Lecture Notes

2. Overview of Techniques for Solving Integer Programs Enumeration Techniques – Complete Enumeration list all “solutions” and choose the best – Branch and Bound Implicitly search all solutions, but cleverly eliminate the vast majority before they are even searched – Implicit Enumeration Branch and Bound applied to binary variables Cutting Plane Techniques – Use LP to solve integer programs by adding constraints to eliminate the fractional solutions.

3. Capital Budgeting Example Investment budget = $14,000 Investment Cash Required (1000s) NPV added (1000s) maximize subject to x j binary for j = 1 to 6

4. Complete Enumeration Systematically considers all possible values of the decision variables. – If there are n binary variables, there are 2 n different ways. Usual idea: iteratively break the problem in two. At the first iteration, we consider separately the case that x 1 = 0 and x 1 = 1.

5. An Enumeration Tree Original problem

6. On complete enumeration Suppose that we could evaluate 1 billion solutions per second. Let n = number of binary variables Solutions times – n = 30, 1 second – n = 40, 17 minutes – n = 50 11.6 days – n = 60 31 years

7. On complete enumeration Suppose that we could evaluate 1 trillion solutions per second, and instantaneously eliminate 99.9999999% of all solutions as not worth considering Let n = number of binary variables Solutions times – n = 70, 1 second – n = 80, 17 minutes – n = 90 11.6 days – n = 100 31 years

8. Branch and Bound The essential idea: search the enumeration tree, but at each node 1. Solve the linear program at the node 2. Eliminate the subtree (fathom it) if 1. The solution is integer (there is no need to go further) or 2. The best solution in the subtree cannot be as good as the best available solution (the incumbent) or 3. There is no feasible solution

9. Branch and Bound Node 1 is the original LP Relaxation maximize subject to Solution at node 1: The IP cannot have value higher than 44 3/7.

10. Branch and Bound Node 2 is the original LP Relaxation plus the constraint x 1 = 0. maximize subject to Solution at node 2:

11. Branch and Bound Node 3 is the original LP Relaxation plus the constraint x 1 = 1. The solution at node 1 was Note: it was the best solution with no constraint on x 1. So, it is also the solution for node 3. (If you add a constraint, and the old optimal solution is feasible, then it is still optimal.)

12. Branch and Bound Node 4 is the original LP Relaxation plus the constraints x 1 = 0, x 2 = 0. Solution at node 4: 0 0 1 0 1 1 z = 42 Our first incumbent solution! No further searching from node 4 because there cannot be a better integer solution.

13. Branch and Bound The incumbent solution has value 42 We next solved the LP’s associated with nodes 5, 6, and 7 No new integer solutions were found.

14. Branch and Bound The incumbent solution has value 42 We next solved the LP’s associated with nodes 8 -13

15. Summary so far We have solved 13 different linear programs so far. – One integer solution found – One subtree fathomed (pruned) because the solution was integer (node 4) – One subtree fathomed because the solution was infeasible (node 13) – No subtrees fathomed because of the bound

16. Branch and Bound The incumbent solution has value 42 We next solved the LP’s associated with the next nodes. We can fathom the node with z = 42.66. Why?

17. Getting a better bound The bound at each node is obtained by solving an LP. But we know the best integer solution has an integer objective value. If the best integer valued solution for a node is at most 42.66, then we know the best bound is at most 42. Other bounds can also be rounded down.

18. Branch and Bound The incumbent solution has value 42

19. Branch and Bound The incumbent solution has value 42

20. Branch and Bound The incumbent solution has value 42 We found a new incumbent solution!

21. Branch and Bound The incumbent solution has value 43 We found a new incumbent solution!

22. Branch and Bound The incumbent solution has value 43 If we had found this incumbent earlier, we could have saved some searching.

23. Finishing Up The incumbent solution has value 43

24. Lessons Learned Branch and Bound can speed up the search – Only 25 nodes (linear programs) were evaluated – Other nodes were fathomed Obtaining a good incumbent earlier can be valuable – only 19 nodes would have been evaluated. Solve linear programs faster, because we start with an excellent or optimal solution – uses a technique called the dual simplex method Obtaining better bounds can be valuable. – We sometimes use properties that are obvious to us, such as the fact that integer solutions have integer solution values

25. Branch and Bound Notation: – z* = optimal integer solution value – Subdivision: a node of the B&B Tree – Incumbent: the best solution on hand – z I : value of the incumbent – z LP : value of the LP relaxation of the current node – LIST: the collection of active (not fathomed) nodes – Children of a node: the two problems created for a node, e.g., by saying x j = 1 or x j = 0. Initialization: LIST = {original problem} Incumbent: = ∅ z I = -∞

26. Branch and Bound Algorithm INITIALIZE SELECT: If LIST = ∅, then the Incumbent is optimal if it exists, and the problem is infeasible if no incumbent exists; else, let S be a subdivision from LIST. Let x LP be the optimal solution to S Let z LP = its objective value CASE 1. z LP = -∞ (the LP is infeasible) Remove S from LIST (fathom it) Return to SELECT

27. CASE 2. -∞ < z LP <= z I. That is, the LP is dominated by the incumbent) Then remove S from LIST (fathom it) Return to SELECT Branch and Bound Algorithm INITIALIZE SELECT: If LIST = ∅, then the Incumbent is optimal if it exists, and the problem is infeasible if no incumbent exists; else, let S be a subdivision from LIST. Let x LP be the optimal solution to S Let z LP = its objective value

28. CASE 2. -∞ < z LP <= z I. That is, the LP is dominated by the incumbent) Then remove S from LIST (fathom it) Return to SELECT Branch and Bound Algorithm INITIALIZE SELECT: If LIST = ∅, then the Incumbent is optimal if it exists, and the problem is infeasible if no incumbent exists; else, let S be a subdivision from LIST. Let x LP be the optimal solution to S Let z LP = its objective value

29. CASE 3. z I < z LP and x LP is integral. That is, the LP solution is integral and dominates the incumbent. Then Incumbent := x LP ; z I := Z lp Remove S from LIST (fathomed by integrality) Return to SELECT Branch and Bound Algorithm INITIALIZE SELECT: If LIST = ∅, then the Incumbent is optimal if it exists, and the problem is infeasible if no incumbent exists; else, let S be a subdivision from LIST. Let x LP be the optimal solution to S Let z LP = its objective value

30. CASE 4. z I < z LP and x LP is not integral. There is not enough information to fathom S Remove S from LIST Add the children of S to LIST Return to SELECT Branch and Bound Algorithm INITIALIZE SELECT: If LIST = ∅, then the Incumbent is optimal if it exists, and the problem is infeasible if no incumbent exists; else, let S be a subdivision from LIST. Let x LP be the optimal solution to S Let z LP = its objective value

31. Different Selection Rules are Possible Rule of Thumb 1: Don’t let LIST get too big (the solutions must be stored). So, prefer nodes that are further down in the tree. Rule of Thumb 2: Pick a node of LIST that is likely to lead to an improved incumbent. Sometimes special heuristics are used to come up with a good incumbent.

32. Branching One does not have to have the B&B tree be symmetric, and one does not select subtrees by considering variables in order. Choosing how to branch so as to reduce running time is largely “art” and based on experience.

33. Different Branching Rules are Possible Branching: determining children for a node. There are many choices. Rule of thumb 1: if it appears clear that x j = 1 in an optimal solution, it is often good to branch on x j = 0 vs x j = 1. – The hope is that a subdivision with x j = 0 can be pruned. Rule of thumb 2: branching on important variables is worthwhile – e.g., in the location problem, branch on the plant location variables first

34. Different Bounding Techniques are Possible We use the bound obtained by dropping the integrality constraints (LP relaxation). There are other choices. Key tradeoff for bounds: time to obtain a bound vs quality of the bound. If one can obtain a bound much quicker, sometimes we would be willing to get a bound that is worse It usually is worthwhile to get a bound that is better, so long as it doesn’t take too long (see next lecture)

35. What if the variables are general integer variables? One can choose children as follows: – child 1: x 1 ≤ 3 (or x j ≤ k) – child 2 x 1 ≥ 4 (or x j ≥ k+1) How would one choose the variable j and the value k – A common choice would be to take a fractional value from x LP. e.g., if x 7 = 5.62, then we may branch on x 7 ≤ 5 and x 7 ≥ 6. – Other choices are also possible.

36. Summary Branch and Bound is the standard way of solving IPs to optimality. There is art to making it work well in practice. Much of the art is built into state-of-the-art solvers such as CPLEX.

37. A bad example for implicit enumeration maximize subject to xi ∈ {0,1} for i = 1 to 100. Why is this a bad example? What would happen if we used branch and bound, as described earlier?